Skip to content

Add /notes: dated, data-linked board notes - #162

Merged
MaxGhenis merged 10 commits into
mainfrom
notes-route
Sep 5, 2026
Merged

Add /notes: dated, data-linked board notes#162
MaxGhenis merged 10 commits into
mainfrom
notes-route

Conversation

@MaxGhenis

Copy link
Copy Markdown
Contributor

Adds /notes: dated records of board changes and findings, each a few paragraphs whose every number is a placeholder filled from a facts block that tests/test_notes.py recomputes from the frozen snapshot under paper/snapshot/20260501/. Two notes:

  • 2026-09-01, Claude Fable 5.1 added. Board row (86.3, PolicyBench v2: AI tax/benefit benchmark #2 of 33, 1,984/1,984 parsed), the JSON transport, the tool_choice auto sensitivity (87.5) with the like-for-like Claude Fable 5 comparison (86.9; board row 79.9 ran chunked under a forced tool call), and the v2 board condition: move every model to tool_choice auto #139 plan. Data: the release, the sensitivity note, the pinned evidence file, the serving-configuration table, the model page.
  • 2026-09-03, Six SNAP households the top three models deny. 20 of 100 households qualify; GPT-5.6 Sol, Claude Fable 5.1, and Kimi K3 each predict $0 for the same six (scenario_027/030/045/073/108/112, reference $287.68 each, the minimum allotment); five qualify through broad-based categorical eligibility, one passes the federal tests; 10 of 20 eligible are categorical-only (6 income, 4 assets); explanation mention counts by regex. Scenario ids deep-link into the explorer. Data: release assets, frozen files, the pathway recompute, policyengine-us parameter trees.

scripts/snap_pathways.py recomputes each household's SNAP tests (meets_snap_gross_income_test, meets_snap_net_income_test, meets_snap_asset_test, is_tanf_non_cash_eligible, is_snap_eligible) with the same situation builder ground_truth.py uses and writes notes/data/snap_pathways_20260901.csv plus a provenance sidecar (policyengine-us 1.723.0; recomputed SNAP agrees with the frozen 1.755.4 references for all 100). Pathways: ordinary 10, categorical_income 5, categorical_both 1, categorical_assets 4 (NJ, NC, VA, PA), ineligible 80.

Also: SiteHeader takes actionLinks (Paper and Notes everywhere the Paper link appeared), /notes and both note URLs in the sitemap, .gitignore admits notes/data/.

Based on add-fable-5-1 because the notes cite the 33-model snapshot that #160 lands; retargets to main when #160 merges. Draft until then.

Verification on the rebased head: ruff clean; pytest 639 passed, 5 skipped; app lint clean, 101 vitest, next build prerenders /notes and both note routes.

@vercel

vercel Bot commented Sep 3, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
policybench-site Ready Ready Preview Sep 5, 2026 4:56pm UTC

Request Review

MaxGhenis and others added 9 commits September 5, 2026 12:48
- New note, 2026-09-05: GPT-6 Astra debuts second at 88.0% (#2 of 39).
  Row by row against GPT-5.6 Sol on the 1,973 scored outputs: both right
  on 1,816, both wrong on 95, 33 Astra-only misses against 29 Sol-only.
  Astra's solo misses cluster in two invented rules (an employer-sponsored
  insurance premium netted from wages, 13 rows; a disability flag read as
  Medicare eligibility, 10 rows). The row list with both predictions, the
  reference and the judge's note is committed (scripts/astra_vs_sol_rows.py
  regenerates it) and the test recomputes every fact from the frozen payload
  and annotations.
- Notes keep the release their facts were checked against. The renderer
  shows "numbers from release X (board snapshot Y)" under the date; the
  test recomputes facts for notes on the frozen release and, for notes on
  a superseded release, checks the release is one the repository published.
  The frozen payload is read through snapshot_payload (gzip).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…elease line

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
@MaxGhenis
MaxGhenis merged commit 2b8dffd into main Sep 5, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant